CPO @_doubleAI_. Real Madrid socio. Tech geek. Sports fanatic
Israel
Joined January 2011
- Tweets497
- Following190
- Followers153
- Likes259
Exceptional work by @_doubleAI_ own @AlonVin and team showcasing how AI can help AI train and rearchitect itself and exit the known territory of transformers!
1/
What if an LLM could not just edit a model's weights, but re-architect it as it trains?
We tried just that. The resulting architectures exit the known territory of transformers: the coding agents spliced code directly into the network.
On nanochat, the loss fell below ordinary training after each new code layer was introduced. On algorithmic reasoning, accuracy rose from 55% to 84%. We even let DeepSeek-V4.1-Flash act as the agent modifying itself, raising its AIME and HMMT 2026 score from 64% to 71%.
To do all this, the agents built a set of tools we’ve never seen before, which we call mechanistic intervention!
Blog: doubleai.s.gy/llm-redesign
3
186
Edik Mitelman retweeted
In Disney’s Fantasia (1940), Mickey Mouse enchants a broom to carry water.
The broom does exactly what it’s told. But things quickly spiral out of control.
This 85-year-old story captures one of the deepest unsolved problems in AI: alignment. 🧵
ALT Mickey is surrounded by brooms carrying water
Edik Mitelman retweeted
Two must-have properties of a Code Execution Model: perfect verification and online learning. I'm especially fond of online learning, particularly when it comes with vanishing regret 😉
Edik Mitelman retweeted
What if code could think while it runs?
Introducing Maxwell: the world's first Code Execution Model.
Maxwell lives inside programs as they run, and improves them! It's an entirely new class of model, sitting between code and hardware. No humans in the loop.
IMO Maxwell is the most direct way to achieve superoptimisation at scale.
Edik Mitelman retweeted
@_doubleAI_ debuted Maxwell: AI that makes code faster while it runs, with a mathematical proof behind every optimization. A new category we call Code Execution Models (CEMs). Rigorous verification is what makes it work..
Meet Maxwell: heymaxwell.ai
@WSJ: tinyurl.com/bjuknyrh
Launching a new product is hard.
Launching a new model is harder.
Launching a new category in AI is...what we do at @_doubleAI_ !
Excited, thrilled, and genuinely proud to announce that we are launching Code Execution Models (CEMs) today.
AI is used extensively in how humans interact with code. Writing, debugging, reviewing, etc.
We are introducing the first-ever AI that lives between code and silicon and helps execute it. Our first CEM, Maxwell, lives inside the machine, understands and reasons about the running code, and improves it while it runs.
To do this, Maxwell can't make mistakes because there's no human in the loop to catch them.
Our breakthroughs in formal verification enable CEMs to reshape algorithms and instructions as they go, since any accepted improvement will be precisely semantically equivalent to the code that preceded it.
Maxwell's implications for hardware utilization and superoptimization are huge.
We believe it is the key to unlocking 100% of compute.
Welcome to the world, Maxwell. Go do your magic!
To read more and sign up for early access, the link is in the first comment.
2
7
142
Edik Mitelman retweeted
Our breakthrough on the Moore girth problem raises a larger question: when will AI become a great scientist?
In this lecture, I discuss what made this result possible—and why verification remains a fundamental barrier across science.
Watch the lecture youtube.com/watch?v=AIMnYEHj…
Edik Mitelman retweeted
1/ Double Agent, our agentic system, has achieved a breakthrough result on one of the most central problems in extremal graph theory. It constructed explicit regular graphs with girth (8/5) log n, the first improvement in 38 years, and a leap towards the theoretical maximum.
Edik Mitelman retweeted
Can AI reach superoptimisation across the internet, at scale? What if every single workload you ran was tailored to your exact hardware configuration?
Gal Beniamini (@laginimaineb) is speaking about it at @NVIDIAGTC Berlin, Thu 22 Oct, 15:00 CEST.
Edik Mitelman retweeted
Another research from @_doubleAI_
Most RL objectives for language models maximize average reward per generation. They do not explicitly reward maintaining several distinct, high-quality solutions, so output diversity can shrink during training—limiting the gains from best-of-k sampling.
ArgMaxRL directly optimizes the expected quality of the best among k samples for continuous rewards. This better aligns training with test-time scaling: generate multiple candidates, verify them, and select the best.
The result could be models that preserve useful diversity and benefit more from additional inference compute.
doubleai.com/research/argmax…
Edik Mitelman retweeted
1/4
On a hard problem, an AI's first answer is rarely its best. The gains come from generating many tries and keeping the best. But standard RL collapses a model onto one answer, so more samples don't help. ArgMaxRL fixes that for continuous rewards, not just pass-or-fail.
Edik Mitelman retweeted
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.
This culminated in the third one taking over part of OpenAI itself.
All this happened while humans remained more-or-less in the dark about the scope of the conspiracy.
I’ve spent the last three days reading through these reports and trying to understand exactly what happened.
Here is my attempt to tell the whole story in plain English:
dwarkesh.com/p/openai-huggin…
Everyone spent yesterday arguing about whether Jalapeño really beats Blackwell. The more interesting sentence is four paragraphs from the bottom of OpenAI's post.
On selected GPT-OSS attention and MoE blocks, AI-generated kernel implementations ran 1.5 to 1.8 times faster than the ones their human experts wrote. OpenAI says plainly that this applies to the selected blocks and not the full model.
This is a claim about software, not silicon. A first-generation ASIC brought three unplanned open-weight models to production performance in two months, on a brand-new instruction set, with a coding agent doing the kernel work. The scarce thing was never the ability to write a fast kernel. It is the willingness to write ten thousand of them and prove each one still computes the same answer.
2
54
Chip design has no “patch it next week.” Almost correct can mean years and millions lost, or worse.
On @nvidia CVDP, @_doubleAI_ solved 98.5% of tasks and set a record 82.5% area-optimization score, beating @claudeai on Fable at it. Every accepted solution was formally verified, and 100% were correct.
That's the same self-improving automated AI system we had improve Pandas, optimize quantum circuits, and write top-performing CUDA kernels.
These capabilities are coming to engineers everywhere very soon. Follow @_doubleAI_ and stay tuned.
Details: doubleai.com/research/warpsp…
1
83
Performance engineering with correctness doesn't apply only to code. It is very relevant to many fields. So after pointing our @_doubleAI_ expert at pandas, we tasked it to do quantum work. As always, it was very creative and bold, found new solutions, and beat Google and the field by a landslide. Read the blog post for details, and stay tuned for our first product release soon.
1/ No one is breaking Bitcoin any time soon.
But the projected cost of the attack just dropped. WarpSpeed, our AI expert, found a circuit 2.5x more efficient than the one Google's quantum team published in March.
2
77
Why not just point Claude Code at your own code and do this yourself? We tried. Ultracode, hundreds of subagents, 1k+ USD in tokens. 1.3x, and one query in ten came out slower. An agent that can't prove it didn't change your results has to play defense.
We @_doubleAI_ can play offense because we prove correctness, and that's the key.
.@_doubleAI_ rewrote pandas' kernels from scratch, specialized for modern hardware.
The result is a drop-in replacement — same API, same semantics — that runs 38.4x faster on PDS-H. Ten minutes of work becomes seventeen seconds. No code changes.
2
50